Generative AI

A type of AI that can create content

How It Works

Generative AI uses patterns it has learned from existing data to produce new outputs.

Model Types

There are several techniques or model types that fall under the umbrella of generative AI:

• Large language models (LLMs) - E.G. ChatGPT
• Small language models (SLM) - E.G. AI used in IoT or mobile apps
• Generative adversarial networks (GANs) - two AIs working against each other to create media such as images or video, a generator that creates and a discriminator that judges and provides feedback. E.G. Deepfakes
• Transformers - advanced models that understand context in sequences of data such as words in a sentence or pixels in an image. Transformers understand the big picture by paying attention to relationships between all parts of the input. E.G. detecting patterns in logs or summarizing a PDF

How It Is Trained

1. Data collection - the model needs a very large dataset for training
2. Pretraining - the model learns basic patterns from the datasets
   • Unsupervised learning which is training without labels to identify patterns
3. Training / alignment training
   • Supervised learning which is training with labelled data
   • Reinforcement learning with human feedback (RLHF) where the model understands the task but needs human guidance to align it with what is correct
4. Fine-tuning
   • Specializes the model for a specific task
5. Optimization
   • Epochs - each complete pass through the training dataset
   • Pruning - removing unneeded parts of the model to increase efficiency
6. Continuous learning
   • Post deployment
   • Fine tuning based internal and external reports
   • Patching to address hallucinations, bias, drift or security vulnerabilities